CommunityJul 4, 20262026-07-044 minVerified
Why Alignment Shapes LLM Behavior More Than Personality
Apologies, refusals, and sycophancy in LLMs are shaped more by alignment, rewards, and prompting than personality.
SourceJul 4, 20262026-07-044 minVerified
Five-Modal MKGR for Cold-Start PPI Prediction
MKGR combines one sequence modality and four knowledge graphs to improve cold-start PPI prediction over prior baselines.
SourceJul 4, 20262026-07-043 minVerified
Medical AI Beyond Tests to Clinical Reasoning
As multiple-choice medical benchmarks saturate, open-ended clinical reasoning and safety are becoming key measures.
SourceJul 4, 20262026-07-044 minVerified
Open-Weight LLM Safety Beyond Release-Time Alignment
Open-weight LLM safety should be judged not only at release, but by how easily fine-tuning can weaken safeguards later.
SourceJul 4, 20262026-07-043 minVerified
PACE Tests Cheap Proxies For Agent Benchmark Performance
PACE examines whether low-cost non-agent benchmarks can predict expensive agent benchmark performance.
SourceJul 4, 20262026-07-044 minVerified
ReContext Makes Long Context Actually Usable in Reasoning
ReContext highlights that long-context value depends on reusing evidence already in the prompt, not just larger windows.
SourceJul 4, 20262026-07-044 minVerified
Workflow Agents for Verifiable Scientific Paper Reproduction
Why scientific ML paper reproduction needs workflow, progress tracking, and evidence-claim matching beyond code generation.
SourceJul 3, 20262026-07-035 minVerified
Automating Agent Safety Testing With Evidence-Based Verification
A summary of arXiv 2607.01793 on automating agent safety testing from risk discovery to evidence-grounded verification.
SourceJul 3, 20262026-07-034 minVerified
Combining RLVR and Human Demonstrations for Better LMs
A paper on combining RLVR with human demonstrations to train style, structure, and diversity beyond verifiable rewards.
CommunityJul 3, 20262026-07-035 minVerified
How To Compare Code Models Beyond Benchmark Scores
Code model evaluation should weigh real task success, retries, latency, and token cost, not benchmark scores alone.
SourceJul 3, 20262026-07-034 minVerified
DiscoLoop Tests Internal Multi-Hop Reasoning In One Pass
DiscoLoop explores multi-hop reasoning inside a single forward pass without relying on long external CoT tokens.
SourceJul 3, 20262026-07-034 minVerified
Can Multimodal AI Improve Rail Crossing Safety Assessment
Examines whether combining rail crossing images with accident records improves safety assessment and what validation matters.
SourceJul 3, 20262026-07-034 minVerified
OCB Tests Native Office Understanding Beyond PDF QA
OCB evaluates native Office file understanding, revealing document AI limits beyond PDF-based QA.
CommunityJul 2, 20262026-07-024 minVerified
AI Data Centers Depend on Power and Cooling
AI data center risks hinge less on hype than on grid connection, cooling design, water tracking, and permitting.
SourceJul 2, 20262026-07-023 minVerified
DART-VLN Improves Discrete VLN Without Retraining at Test Time
DART-VLN targets stale memory reads and local backtracking in discrete VLN using training-free test-time control.
SourceJul 2, 20262026-07-025 minVerified
Memory-Native NTN Design for Remote Robotics Missions
Examines why remote robots on NTN need memory-based communication that uses past link states and task context.
SourceJul 2, 20262026-07-024 minVerified
Rethinking Human Oversight for High-Risk Autonomous AI
A look at high-risk AI oversight through the humans-as-handlers approach, focusing on intervention, accountability, and trust.
CommunityJul 2, 20262026-07-024 minVerified
How Safety Holds Up Across Long AI Conversations
Examines whether AI safety remains consistent in long conversations and highlights gaps in session-level evaluation.
CommunityJul 1, 20262026-07-015 minVerified
AI Employment Narrative Shifts From Loss to Redesign
Examines whether AI eliminates jobs or redesigns tasks, and why this shift matters for hiring, reskilling, and productivity.
SourceJul 1, 20262026-07-015 minVerified
Design Axes for Agentic Orchestration in Enterprises
A practical guide to balancing agent autonomy, traceability, and control in enterprise orchestration design.
SourceJul 1, 20262026-07-014 minVerified
Emergent Misalignment Depends on More Than Training Data
EM may depend on optimizers and batch settings, making finetuning recipes part of safety evaluation, not just data.
SourceJul 1, 20262026-07-015 minVerified
How Agents Should Help Users Form Preferences
Why AI agents must move beyond preference elicitation to support preference formation, with evaluation and safety in view.
SourceJul 1, 20262026-07-014 minVerified
LLM Agents Split Research Ideas And Verification
How a dual-agent LLM pipeline separates proposing tighter relaxations from verification in automated research.
SourceJul 1, 20262026-07-014 minVerified
Rethinking AI Tutors Beyond Cloud Chatbots in Education
In education, AI design matters more than raw performance, with student privacy, data minimization, and teacher control at stake.